JPH0451037B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a voice recognition method for causing a machine to recognize a human voice.

BACKGROUND OF THE INVENTION In recent years, voice recognition technology has been actively developed and commercialized, but most of these are for specific speakers whose voices are recognized only by those who have registered their voices.
Devices for specific speakers require time and effort to register the words to be recognized in the device in advance, which puts a heavy burden on the user unless the device is used continuously for a long time. In response to this, research has recently been actively conducted on recognition technology for non-specific speakers that is easy to use and does not require voice registration.

Generally speaking, the speech recognition method performs pattern matching between the input speech and the standard speech stored in the dictionary (these are parameterized), and selects the one in the dictionary with the highest degree of similarity. This means that the voice is output as the recognition result. In this case, there is no problem if the input voice and the voice in the dictionary are physically exactly the same, but in general, even if the input voice is the same voice, different people or different ways of saying it may cause it to be completely different. It won't be the same.

Physically, differences between people and differences in the way they speak are expressed as differences in spectral features and differences in temporal features. In other words, the shape of the articulatory organs (mouth, tongue, throat, etc.) differs from person to person, so the spectral shape of the same word will differ between different people.
Furthermore, the temporal characteristics differ depending on whether the voice is spoken quickly or slowly.

Speaker-independent recognition techniques require such spectra and their temporal variations to be normalized and compared to standard patterns.

The present applicant has already filed a patent application for a method that uses parameter time series information and a statistical distance measure in combination (Japanese Patent Application No. 60-29547) as an effective method for speech recognition of unspecified speakers. The method will be explained below.

FIG. 10 is a functional block diagram showing an implementation of the speech recognition method previously proposed by the applicant.

In the figure, 1 is an AD conversion unit that converts input audio into a digital signal, 2 is an acoustic analysis unit that analyzes audio for each analysis section (frame) and obtains spectrum information, 3 is a feature parameter extraction unit that obtains feature parameters, and 4 5 is a speech interval detection unit that detects the start frame and the end frame, 5 is a time axis normalization unit that expands and contracts the word length, 6 is a distance calculation unit that calculates the similarity between the input pattern and the standard pattern, and 7 is created in advance. This is a standard pattern storage unit that stores standard patterns that have been created. The operation of the above configuration will be explained below.

The input audio is converted into a 12-bit digital signal by the AD converter 1. Sampling frequency is 8KHz
It is. The acoustic analysis section 2 performs LPC analysis using the autocorrelation method for each frame (10 msec). The order of analysis is 10th, and the linear prediction coefficient is 0,
Find 〓1,〓2...〓10. In addition, the audio power W _p for each frame is also determined here. The feature parameter extraction unit 3 uses the linear prediction coefficients to obtain LPC cepstral coefficients C ₁ to C _p (p is the truncation order) and normalized logarithmic residual power C _p . In addition, LPC
The analysis and method for extracting LPC cepstrum coefficients are described in detail in, for example, "Linear Prediction of Speech" by JD Markel and AH Gray, translated by Hisaki Suzuki, so the explanation will be omitted here. In addition, the feature parameter extraction unit 3 obtains the logarithmic power LW _p using the following equation.

LW _p = 10log10W _p (Formula 1) The speech interval detection unit 4 compares LW _p obtained by (Formula 1) with the threshold θ _S , and if a frame with LW _p > θ _s continues for l _s frames or more, the first Let the frame be the starting frame _Fs of the voice section. Further, after F _s , LW _p is compared with the threshold value θ _e , and if there are consecutive frames with LW _p < θ _e for _LE frames or more, the first frame is set as the end frame F _e of the voice section.
In this way, the period from F _s to F _e is defined as a voice section.
To simplify the explanation, let us consider F _s as the first frame and set the frame number as (1, 2,...
j,...J). However, J=F _e −F _s +1.

The time axis normalization unit 5 linearly expands and contracts the word length by dividing it into I-frame lengths. The i-th frame after expansion and contraction and the j-th frame of the input audio have the relationship expressed by (Equation 2).

i=[J-1/J-1j+J-/J-1+0.5]
(Formula 2) However, [ ] represents the largest integer that does not exceed that number. In the example, I=16.

Next, the feature parameters after expansion and contraction are arranged in chronological order,
Create time series pattern 〓 _x . Feature parameters of the current i-th frame (LPC cepstral coefficients)
When C ^x _i , k (k=0, 1, 2,...P: d pieces), _x becomes the following formula.

〓 _x = (C ^(x) ₁ , 0, C ^(x) ₁ , 1, C ^(x) ₁ , 2...C ^(x) ₁
,p
……C ^(x) _i 0,C ^(x) _i ,1……C ^(x) 〓,0C ^(x) 〓,1
…
C ^(x) 〓,P) (Formula 3) That is, 〓 _x is ・(P+1), that is, ・D
dimensional vector (D is the number of parameters per frame).

The distance calculation unit 6 calculates the degree of similarity between the input pattern _x and the standard pattern of each voice stored in the standard pattern storage unit 7 using a statistical distance measure, and selects the voice with the smallest distance as the recognition result. Output as . The standard pattern corresponding to the k-th voice stored in the standard pattern storage unit 7 is 〓 _k
(average value), and the covariance matrix common to all target speech is 〓, the Mahalanobis distance S _k between the input pattern 〓 _k and the k-th standard pattern is calculated by the following formula.

S _k = (〓 _x −〓 _k ) ^t・〓 ⁻¹・(〓 _x −〓 _k ) (Formula 4) The subscript t represents transposition, and −1 represents an inverse matrix. Expanding (Equation 4), S _k =〓 ^t _x・〓 ^-1・〓 _x −2〓 ^t _k・〓 ^-1・〓 _x +〓 ^t _x・〓 ^-1・〓 _k (Formula 5) (Formula 5 ) is unrelated to n, so it does not need to be taken into consideration when comparing sizes. Therefore, the first
If we remove the terms and replace S _k with D _k , D _k becomes:

D _k = b _k −〓 ^t _k・〓 _x (Formula 6) Where〓 _k = 2〓 ^-1・〓 _k (Formula 7) b _k =〓 ^t/k・〓 ^-1・〓 _k (Formula 8) Dk is calculated for all k (k=1, 2...N), and the speech that minimizes Dk is taken as the recognition result. Here, k is the number of voice standard patterns stored in the standard pattern storage section 7. Actually, the standard pattern is stored as a pair of 〓k and bk, and the number of voices (K types) is stored.

The amount of calculation required for (Equation 6) is I・(P
+1) times and one subtraction, and the feature is that the amount of calculation is extremely small. Practically, I=16, P=4
is sufficient, so the number of product-sum operations is 80 per word.

Next, the standard pattern 〓k, 〓 (actually 〓k,
This section explains how to create (converted to bk).

A standard pattern is created using many data samples for each voice. For each voice, let M be the number of samples used. Apply (Equation 2) to each sample to make the number of frames equal to I.
Find the average value vector for voice k.

〓k=(C ^(k) ₁ , 0, C ^(k) ₁ , 1, C ^(k) ₁ , 2,...C ^(k
) ₁ ,
p……C ^(k) _i ,0,C ^(k) _i ,1……C ^(k) 〓,0,C ^(k)
〓、
1,...C ^(k) 〓,P) (Formula 9) where C ^(k) _i ,n=1/M _M 〓 ^m=1 Ci, ^(k) _o ,m (Formula 10) i=1,2, ...I: I frame n=0, 1, 2, ...P: d pieces Here, Ci, n, m are the m-th sample of audio k and indicate the n-th order cepstral coefficient of the i-th frame. A covariance matrix W ^k of voice k is obtained using the same procedure as for the mean value vector. The covariance matrix 〓 common to all voices is calculated using the following formula.

〓=1/K (〓(1)+〓(2)+…+〓 ^k +…+〓 ^K ) (Formula 11) 〓k,〓 by (Formula 7) and (Formula 8), 〓k,
bk and stored in the standard pattern storage section 7 in advance.

Problems to be Solved by the Invention The problem with this method is that it assumes that the speech interval is uniquely and reliably determined before pattern matching is performed. Actual speech data contains various types of noise, and occurrences at the beginning and end of words are unclear, so there are many cases where it is not possible to accurately determine speech intervals, or where non-speech intervals are mistakenly detected. . If the conventional method is applied to the erroneously misidentified speech section, the recognition rate will naturally drop significantly.

The purpose of this invention is to solve the above problems.
It is an object of the present invention to provide a speech recognition method having a high recognition rate and capable of automatically extracting and recognizing speech from an input signal without requiring the operation of detecting a speech section.

Means for Solving the Problems The present invention achieves the above object, and uses a sufficiently long section including the speech to be recognized and the noises before and after it as an input signal section, and sets a certain time standard in this input signal section. Set a point, use the reference point as an end point, and then
N ₁ frame interval and N ₂ frame interval (N ₁ <
N ₂ ), and considering these as the minimum and maximum values of the voice section, N ₂ −N ₁ +1
For each speech segment candidate, the similarity or distance of each word is determined by matching the speech segment length with the standard pattern of each word while expanding or contracting it to a certain length of time. Scans all input signal sections from the beginning to the end, compares the similarity or distance of all reference point positions to all speech section candidates for each word, and recognizes the word with the maximum similarity or minimum distance. This is what is output as a result.

Effect The present invention performs pattern matching between linearly expanded and contracted input and a standard pattern while shifting one frame at a time for all input signal sections, and automatically selects the audio with the maximum similarity or the minimum distance and its section. Since the method is calculated based on the method, it is no longer necessary to detect speech sections, and speech uttered in a noisy environment can be recognized with high probability.

Examples Examples of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a functional block diagram showing an embodiment of a speech recognition method according to an embodiment of the present invention.

First, the concept of the embodiment will be explained using FIGS. 2 to 4. Even when the same word is uttered, the temporal length of the utterance (voice length) differs depending on the utterance method and differs depending on the person. In the speech recognition method using pattern matching, the length of the input speech is
After normalizing to standard speech, similarity calculation is performed to recognize speech length. FIG. 2 shows how the voice length is normalized. Let N ₁ be the minimum length of input audio, N ₂ be the maximum length, and I be the standard length of audio (standard pattern length), then the second
As shown in the figure, the voice length of length N (N ₁ ≦N≦N ₂ ) is expanded or contracted to be normalized to length I. In Fig. 2, the ends of the audio are made to match and are expanded and contracted. For expansion and contraction, a linear expansion and contraction formula is used, similar to (Formula 2).

i=[I-1/N-1・n+N-I/N-1+0.5
] (Formula 12) When calculating the similarity between an unknown input and a standard pattern, the voice length N of the unknown input is expanded or contracted to the standard pattern length by (Formula 12). FIG. If we take the input length on the horizontal axis and the standard pattern length on the vertical axis, and match the ends, the input audio length is in the range of N ₁ to _{N 2} , so the matching route between the input and the standard pattern is the input axis. A straight line starts from one point within N ₁ ≦N≦N ₂ and ends at P. Therefore, all similarity calculations are performed inside the triangle.

Assume now that there is an unknown input of length N _U and its content is voice k. However, it is assumed that although the terminal end of the unknown input is known, the terminal end is unknown (therefore, N _U is also unknown). When matching this unknown input with the standard pattern S _k of word k, N
is shifted from N ₁ to N ₂ one frame at a time, the time length is expanded or contracted to I using (Equation 12) for each frame, and the degree of similarity between the unknown input parameter and the standard pattern is determined. In this case, the standard pattern is
Since S _k , if the utterance is accurate, the similarity should be maximum at N=N _U. Furthermore, the degree of similarity should be greater for S _k than for any other standard pattern S _k ′. In this way, the starting end of the unknown input is determined (therefore, the voice length is determined) and at the same time voice k can be recognized.

Now, in FIG. 3, the explanation has been made assuming that the end is known, but this method can also be extended to the case where both ends are unknown (that is, when the voice section is unknown). FIG. 4 is an explanatory diagram thereof. In the figure, the horizontal axis (input time axis) coordinate of the terminal point is j. Here, if the position of j coincides with the end of the input voice, the situation is the same as in Figure 3, but this time it is assumed that both end points are unknown, so j does not necessarily coincide with the end of the voice. is not limited. However, if j is a wide range j ₁ ≦j≦
If we scan with _j2 , there will always be a time point j= _j0 where j coincides with the end of the voice. In that case, the starting simple is
It should exist at the point j ₀ −N _U within the range of j ₀ −N ₂ to _{j 0} −N ₁ . Even when scanned in this way, if the uttered word and the standard pattern match, the similarity when the starting point is j ₀ −N _U and the ending point is j ₀ is different from any other j and It is larger than the combination of N. Furthermore, this degree of similarity is greater than the degree of similarity with respect to other standard patterns. Therefore, at the same time as the recognition result is obtained, the start and end points of the voice are determined.

In this manner, the method shown in FIG. 4 can extract and recognize the portion most similar to the standard pattern from a signal containing a mixture of noise and voice. Therefore, there is no need for a commonly used complicated speech segment detection procedure, and the speech segment is output as a result together with the recognized speech.

As described below, the degree of similarity is calculated using a time-series pattern of feature parameters using a statistical distance measure (distance based on posterior probability).

The number of feature parameters per frame is D
Then, the time series pattern of I frame is D _X I
It becomes a vector of dimensions. Now, if the parameter of the i-th frame of the unknown input is 〓 _i and the component of the i-th frame of the standard pattern of word k is a ^k _i , then 〓 _i = (x ₁ , i, x ₂ , i, ... x _d , i,...x _D ,
i) (Formula 13) 〓 ^k _i = (a ^k ₁ , i, ... a ^k ₂ , i, ... a ^k _d , i, ... a ^k _D ,
i) (Formula 14) Let the time series patterns be 〓 and 〓 _k , respectively. 〓 = (〓 ₁ , 〓 ₂ , ..., 〓 _i , ..., 〓 _I ) (Formula 15) 〓 _k = (〓 ^k ₁ , 〓 ^k ₂ ,...〓 ^k _i , ...,〓 ^k _I ) (Equation 16). If the similarity for word k is L _k , then L _k =B _k −〓 ^t _k・〓 (Formula 17) =B _k − _I 〓 ⁱ⁼¹ (〓 ^k _i ) ^t・〓 _i (Formula 18) =B _k − _I 〓 ⁱ⁼¹ ( _D 〓 ^z=1 a ^k _d , i x _d , i) (Formula 19) where 〓 _k , B _k is a standard pattern of word k.

〓 _k = 2〓 ^-1 〓(〓 _k −〓 _e ) (Formula 20) B _k =〓 ^t _k・〓 ^-1 〓・〓 _k −〓 ^t _x・〓 ^-1 _a・〓 _e
(Equation 21) However, 〓 _k is the average value vector of word k, and 〓 _e is the average value vector of surrounding information of all words.
Also, 〓 _a is a covariance matrix, which can be created using the covariance matrix 〓 _k of each word and the covariance matrix 〓 _e of surrounding information. _a = _k 〓 ^k=1 (〓W _k + _e )/(K+1) (Formula 22) k is the type of word.

〓 _e and 〓 _e are created as follows using many samples belonging to each word. As shown in FIG. 5, a plurality of sections (section length is I frame) are set for the voice and its surrounding sections by shifting them one frame at a time. This operation is performed on many samples of each word, and an average value vector 〓 _e and a covariance matrix 〓 _e of the parameters of those intervals are created.

Since (Formula 17) has the same form as (Formula 6), the amount of calculation required for calculating the similarity is the same as in the conventional example. Equations for creating standard patterns ((Equation 7), (Equation 8) and (Equation
20) and (Equation 21)) are the only differences. A feature of the present invention is that surrounding information is incorporated into the standard pattern as 〓 _e and 〓 _e . In this way, (formula
17) is a distance based on pseudo posterior probability.

In FIG. 1, 10 is an AD converter that converts the input signal into a digital signal, 11 is an acoustic analyzer that analyzes each audio analysis section (frame), and 12 is a feature parameter extractor.
Outputs LPC cepstrum coefficients (C ₀ to C ₅ ) every frame (10 msec). The output of the feature parameter extraction unit 12 corresponds to ≦ of (Equation 13) (therefore, D=6). The functions of blocks 10-12 are the same as those of blocks 1-3 in FIG. In addition to LPC cepstrum coefficients, feature parameters include autocorrelation coefficients, PARCOR coefficients, and bandpass filter outputs.

The functions of each block will be explained below with reference to the flowchart of FIG. The frame synchronization signal generator 13 generates a synchronization signal for each frame. Let the frame number be j, and calculate the similarity in a sufficiently wide interval j ₁ ≦j≦j ₂ that includes the input voice. The following operations are performed in one frame period.

The standard pattern selection unit 18 selects each voice (word here) to be recognized (the number of words is K). For the selected standard pattern, the section candidate setting unit 15 sets the minimum speech section length N ₁ (k) and maximum speech section length N ₂ (k) for each word. Then, the section length N (N ₁ (k)≦N<N ₂ (k))
, the unknown input parameters obtained by the feature parameter extraction unit 12 are arranged for the time period of j−n to j frames to create a time series of input parameters,
In the time axis normalization unit 14, the time of the time series parameter is expanded or contracted to an I frame using (Equation 12),
Obtain a parameter series corresponding to (Equation 15). The similarity calculation unit 16 uses this parameter series and the standard pattern storage unit 1 selected by the standard pattern selection unit 18.
Between standard pattern 〓 _k and B _k in 7, (Equation 17)
The similarity L _k (N) is calculated using . The similarity comparison unit 20 compares L _k (N) with the maximum similarity value (minimum distance Lmin) stored in the primary memory 19 up to this point, and if L _k (N) < Lmin, then Lmin
Replace with L _k (N) and set k at that time to k^ to 1
The next memory 19 is updated, and if L _k n≧Lmin, the contents of the primary memory 19 are not updated.

This series of operations is performed on K standard patterns during one frame, _N2 (k) _-N1 (k)+1 times for each standard pattern. Further, this is performed for the period of frames j ₁ to j ₂ . The recognition result is k^ at the time when _j2 frames are reached, and the similarity value at that time is Lmin
It is. Also, the frame j^ at the time when the maximum similarity is obtained
If the interval length N^ at that time is stored in the primary memory 19, the voice interval can be obtained as a result using these.

As described above, in this embodiment, the interval j ₁ to j ₂ is
As long as the area is set wide enough to accommodate the voice, the voice can be recognized without the need for voice section detection. The first embodiment shown in FIG. 1 is easy to understand and is useful for explaining the method, and it is of course possible to implement it in this manner. However, when attempting to achieve real-time implementation, there is a problem in that the amount of calculation is too large. The reason for this is that (Equation 17) is properly calculated for all the sections set by the section candidate setting section 15.

The second embodiment described below is a more practical method that reduces the amount of calculation. First, I will explain the principle.

To obtain the recognition result, k=k^, which minimizes L _k , can be found in similarity calculation formula (18). That is, minL _k = min{B _k − _I 〓 ⁱ⁼¹ (〓 ^k _i ) ^t・〓 _i } =B _k −max{ _I 〓 ⁱ⁼¹ (〓 ^k _i ) ^t・〓 _i } (Equation 23) =B _k −max{ _I 〓 ⁱ⁼¹ l ^k _i (N)} (Formula 24) =B _k −maxM ^k (N) (Formula 25) where l ^k _i (N) = (a ^k _i ) ^t・〓 _i (Formula 26) is the partial similarity between the i-th frame input 〓 _i and the standard pattern k after being time-stretched according to the matching route N. Next, let's consider the meaning of time expansion and contraction. Letting the unknown input vector before time warping be 〓, it is expressed as 〓=(〓 ₁ , 〓 ₂ , ...〓 _o , ...〓 _N ) (Equation 27). n and i are both integers,
They are related by (Equation 12). Therefore (formula
The vector 〓 in 15) is obtained by selecting I frames of frames related by (Formula 12) from among the unknown input vectors 〓 in (Formula 27) and arranging them in temporal order. For convenience, the operation of selecting according to the matching route is expressed by the following equation.

〓 _i ＝〓〓 _i 〓N (Formula 28) Then, the partial similarity (Formula 26) is l ^k _i (N) = ( ^k _i ) ^t・〓〓 _i 〓N (Formula 29) Also, the sum of the partial similarities M ^k (N) is M ^k (N) = _I 〓 ⁱ⁼¹ l ^k _i (N) = _I 〓 ⁱ⁼¹ (〓 ^k _i ) ^t・〓 _i 〓N (Equation 30) That is, (Equation 17) is If the partial similarities l ^k _i (N) have been found in advance, they can be replaced by adding them for I frames according to the relationship in (Equation 12). (Equation 12) uniquely determines the relationship between i and n when N is given, so it can be calculated in advance within the range of N ₁ ≦N < N ₂ and stored in a table or the like.

Next, with reference to FIG. 7, we will consider how to obtain l ^k _i (N). In the figure, point P is the terminal point of the standard pattern and unknown input, and the coordinates of the terminal point of the unknown input are N ₀ . N ₁ and N ₂ are the minimum and maximum lengths of speech as before. Now, assuming that the degree of similarity is determined when the starting point of the unknown input is N, the matching route is the straight line PN. on PN (Equation 12)
The partial similarity l _i (N) at any point (n', i) that satisfies is the product of the vector of the input n' frame and the vector of the i-frame component of the standard pattern 〓 ₁ . Point (n', i) is located on PN at the present time, but since point P shifts with time, it should have been on P'N'O before frame n'. Therefore, at point P, (n′,
Find the partial similarity of i) and accumulate it,
It can be used at point P′. Since (n', i) is an arbitrary point on ΔPN ₂ N ₁ , the same can be said for other points. Thinking like this,
The calculation in each frame can be divided into two parts as follows.

The partial similarity on PN _O is calculated and accumulated in the buffer. (Product-Sum Calculation) The l ^k _i (N) used for the partial similarity sum calculated by (Equation 30) is extracted from the value calculated in the previous frame and stored in the buffer. (Addition operation) FIG. 8 is a block diagram showing the calculation method per frame. In the figure, 30 is l ^k _i
(N _O ), and is prepared for the number of frames (I) of the standard pattern. From the bottom of each product-summer, the input vector of the j-th frame 〓
(j) is input, and the standard pattern is input from the left side. Then, a calculation corresponding to (Equation 29) is performed and l ^k _i (N _O ) is output. The delay buffer 31 is
The calculation results of the product-summer are saved for a period of one frame,
Propagate to the next stage. The number of delay buffers provided for each word is equal to the number of points within ΔPN ₂ N ₀ in FIG. Reference numeral 32 denotes an adder, which calculates the sum of similarities by performing calculations equivalent to (Equation 30). Adder 32
has I input terminals, each of which is connected to the output terminal of the delay buffer according to the matching route defined by (Equation 12). 33 is a comparator, which calculates maxM _k (N). 34 is a subtracter, which calculates the formula (25) to find the minimum value for word k.

The method in the second embodiment has been explained above. FIG. 9 is a functional block diagram showing an implementation of the speech recognition device in the second embodiment. In FIG. 9, blocks having the same numbers as in FIG. 1 have the same functions, so their explanation will be omitted or simplified.

In FIG. 9, an AD converter 10, an acoustic analyzer 11, and a feature parameter extractor 12 digitize input audio and perform LPC analysis to obtain feature parameters (LPC cepstral coefficients) for each frame. The following operations are performed within the period of one frame.

The standard pattern selection unit 18 selects the K standard patterns stored in the standard pattern storage unit 17.
Select one at a time. The partial similarity calculation unit 21 calculates (Equation 29) between the input feature parameter and the selected standard pattern to obtain l ^k _i (N _O ), and stores it in the similarity buffer 22 . The similarity buffer has a capacity to store similarities within ΔPN ₂ N ₀ in FIG. The time expansion/contraction table describes the relationship between n and i defined by (Equation 12) for each input length N (N ₁ ≦N≦N ₂ ). N ₁ and N ₂ differ for each word and are set by the section candidate setting unit 15. The similarity addition unit 23 adds the output of the similarity buffer 22 read at the address specified in the time expansion/contraction table 24 to each of the matching routes N ₁ to _{N 2} to calculate (Equation 30). and calculate the sum of similarities
Find M ^k (N). The similarity comparison unit 20 uses M ^k (N)
and the contents of the primary memory 19 are compared, and only if M ^k (N) is larger, the contents of the primary memory are replaced with M ^k (N). After calculating N=N ₂ (Equation 18)
L _k is obtained by and compared with the previous minimum value stored in the primary memory 19, and the contents of the primary memory 19 are updated only when L _k is small. Then, the standard pattern selection unit 18 selects the next word and performs the same operation. Furthermore, when all words are completed, the frame advances.

If such an operation is performed for the entire target interval (j = j ₁ to j ₂ ), at the end of j = j ₂ frames, the minimum similarity value L^ and the word name at that time will be calculated.
k^ can be obtained as a recognition result.

In the second embodiment, compared to the first embodiment, the number of sum-of-products operations for determining the degree of similarity is significantly smaller. Now, assuming that the number of words K = 10, the standard pattern length I = 16, the average minimum time length N ₁ = 21, the average maximum time length N ₂ = 40, and the number of parameters per frame D = 6, the first embodiment The amount of product-sum calculations in the second embodiment is 19,800 times, whereas it is 960 times in the second embodiment.

Using the method of this example, a total of 330 adult men and women evaluated 10 numeric words generated through telephones, and as a result, an average recognition rate of 93.75% was obtained. This value cannot be said to be low considering that the speech is made under high noise conditions. Furthermore, as a result of analyzing the causes of recognition errors according to this embodiment, it was found that most errors occur because a part of a certain word is recognized as another word. For example, the / ro / part of /Zero/ is mistakenly recognized as /go/.
Therefore, if up to the second candidate are correct, a recognition rate of 97% or higher is obtained. Therefore, it is easily inferred that if a few other methods are used in combination, an even higher recognition rate can be obtained as the first candidate.

Effects of the Invention In summary, the present invention provides a certain temporal reference point in an input signal section that includes the speech to be recognized and the noise before and after the speech, and uses the reference point as an end point and then N ₁
Frame interval and N ₂ frame interval (N ₁ < N ₂ )
Set two intervals, consider these as the minimum and maximum values of the voice interval, and expand or contract the voice interval length to a constant time length for each of N ₂ −N ₁ +1 voice interval candidates. The similarity or distance of each word is determined by matching each word with a standard pattern, and this operation is performed by scanning the reference point from the beginning to the end of the entire input signal section, and all of the reference point positions are This method compares the degree of similarity or distance for each word with respect to candidates for speech sections, and outputs the word with the maximum similarity or the minimum distance as the recognition result.It does not require detection of speech sections and can be used to recognize a mixture of noise and speech. It is possible to extract and recognize only the part corresponding to speech from the signal. Conventionally, complex rules were used to detect speech sections, but even then, when the noise level is high or non-stationary noise is mixed in. However, the present invention simplifies the system by removing the complicated speech interval detection algorithm, and also makes it more suitable for high-noise inputs. A stable recognition rate can be ensured, and the effect is significant.

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram embodying the speech recognition method according to the first embodiment of the present invention, FIGS. 2 to 4 are conceptual diagrams illustrating expansion and contraction of the speech interval length in the same embodiment, and FIG. 5 6 is a conceptual diagram illustrating a standard pattern creation method for surrounding information when creating a standard voice pattern in the same embodiment, FIG. 6 is a flowchart illustrating the processing procedure of the same embodiment, and FIG. FIG. 8 is a block diagram showing the calculation method per frame in the second embodiment, and FIG. FIG. 10 is a functional block diagram showing a conventional speech recognition method. 10...AD conversion section, 11...acoustic analysis section, 1
2... Feature parameter extraction unit, 13... Frame synchronization signal generation unit, 14... Time axis normalization unit, 15
...Section candidate setting unit, 16...Similarity calculation unit, 1
7... Standard pattern storage section, 18... Standard pattern selection section, 19... Primary storage, 20... Similarity comparison section.

Claims (1)

[Claims] 1. A standard pattern for each voice to be recognized is created in advance using data belonging to each voice, data for all voices to be recognized, and surrounding information of the data for all voices. On the other hand, a temporal reference point is set in the unknown input containing the speech to be recognized and its surrounding information, and two time lengths N ₁ and N ₂ (N ₁ < N ₂ ) are set from the reference point as an end point. Set the interval, consider the interval between the reference point and N ₁ to be the minimum value of the voice interval, and the interval between the reference point and N ₂ to be the maximum value of the voice interval, and set the interval between the minimum voice interval and the maximum voice interval. Assuming multiple speech intervals, the length of each assumed speech interval is expanded or contracted to a certain length of time, and the similarity or distance for each voice is determined by comparing each voice with the standard pattern. Memorize the maximum similarity or minimum distance and the standard pattern name in that case for all standard patterns in the speech interval, then shift the reference point in the unknown input by a unit interval, and calculate the new maximum similarity or minimum distance in the same way. , compare the previously stored maximum similarity or minimum distance with the new maximum similarity or minimum distance, memorize the larger similarity or smaller distance and the standard pattern name at that time, and This operation is performed for a sufficiently wide range of unknown input while shifting the reference point by unit time, and when the reference point reaches the final point, the speech corresponding to the memorized standard pattern name is recognized as the result. A speech recognition method characterized by: 2. In advance, find the correspondence between the speech interval length and the temporal position of the standard pattern when the speech interval length is expanded or contracted to a certain time length. On the other hand, when calculating similarity or distance, unknown input and A patent characterized in that the partial similarity or distance of a standard pattern is first determined, and the partial similarity is added to the similarity or distance between an unknown input of an assumed speech interval length and the standard pattern while referring to the correspondence relationship. A speech recognition method according to claim 1. 3. The speech recognition method according to claim 1, wherein the similarity or distance is calculated using a measure based on a posteriori probability. 4 Feature parameters are LPC cepstral coefficients,
2. The speech recognition method according to claim 1, wherein the speech recognition method is either an autocorrelation coefficient or an output of a bandpass filter. 5. A patent characterized in that surrounding information is statistically created from many data samples belonging to all target words, using a speech interval determined by combining ₁ frame exactly near the beginning and ₂ frames exactly near the end. A speech recognition method according to claim 1. 6. The speech recognition method according to claim 1, wherein a standard pattern for a certain speech n is obtained by removing surrounding information from a standard pattern statistically obtained using data belonging to n. 7. The speech recognition method according to claim 1, wherein the formula for calculating the degree of similarity is a linear discriminant function.